Adding upper_face_maps index check, removing plot lower bound, making device agnostic test implementation and correcting docstring mistakes

Author: VijayVignesh1

src/torchmetrics/functional/multimodal/fdd.py

@@ -27,7 +27,8 @@ def upper_face_dynamics_deviation(
 
     The Upper Face Dynamics Deviation (FDD) metric evaluates the quality of facial expressions in the upper
     face region for 3D talking head models. It quantifies the deviation in vertex motion dynamics between the
-    predicted and ground truth sequences by comparing the temporal variation (standard deviation) of per-vertex squared displacements from the neutral template.
+    predicted and ground truth sequences by comparing the temporal variation (standard deviation) of per-vertex
+    squared displacements from the neutral template.
 
     The metric is defined as:
 
@@ -58,7 +59,7 @@ def upper_face_dynamics_deviation(
             If the number of dimensions of `vertices_pred` or `vertices_gt` is not 3.
             If vertex dimensions (V) or coordinate dimensions (3) don't match.
             If ``upper_face_map`` is empty or contains invalid indices.
-            If there are at least two frames to compute face dynamics deviation.
+            If there are fewer than two frames to compute face dynamics deviation.
 
     Example:
         >>> import torch
@@ -90,7 +91,6 @@ def upper_face_dynamics_deviation(
             f"upper_face_map contains invalid vertex indices. Max index {max(upper_face_map)} is larger than "
             f"number of vertices {vertices_pred.shape[1]}."
         )
-
     min_frames = min(vertices_pred.shape[0], vertices_gt.shape[0])
     pred = vertices_pred[:min_frames, upper_face_map, :]  # (T, M, 3)
     gt = vertices_gt[:min_frames, upper_face_map, :]

src/torchmetrics/multimodal/fdd.py

@@ -55,8 +55,6 @@ class UpperFaceDynamicsDeviation(Metric):
         V is the number of vertices, and 3 represents XYZ coordinates.
     - ``target`` (:class:`~torch.Tensor`): Ground truth vertices tensor of shape (T, V, 3) where T is the number of
         frames, V is the number of vertices, and 3 represents XYZ coordinates.
-    - ``upper_face_map`` (:class:`list`): List of vertex indices corresponding to the upper-face region.
-    - ``template`` (:class:`~torch.Tensor`): Template mesh tensor of shape (V, 3) representing the neutral face.
 
     As output of ``forward`` and ``compute``, the metric returns the following output:
 
@@ -89,7 +87,6 @@ class UpperFaceDynamicsDeviation(Metric):
     is_differentiable: bool = True
     higher_is_better: bool = False
     full_state_update: bool = False
-    plot_lower_bound: float = 0.0
 
     vertices_pred_list: List[Tensor]
     vertices_gt_list: List[Tensor]
@@ -108,7 +105,12 @@ def __init__(
             raise ValueError(f"Expected template to have shape (V, 3) but got {template.shape}.")
         if not self.upper_face_map:
             raise ValueError("upper_face_map cannot be empty.")
-
+        if min(self.upper_face_map) < 0 or max(self.upper_face_map) >= self.template.shape[0]:
+            raise ValueError(
+                f"upper_face_map contains invalid vertex indices. "
+                f"Valid indices are between 0 and {self.template.shape[0] - 1}, "
+                f"but got min index {min(self.upper_face_map)}, max index {max(self.upper_face_map)}."
+            )
         self.add_state("vertices_pred_list", default=[], dist_reduce_fx=None)
         self.add_state("vertices_gt_list", default=[], dist_reduce_fx=None)
 

tests/unittests/multimodal/test_fdd.py

@@ -46,9 +46,9 @@ def _generate_vertices(batch_size: int = 1) -> _InputVertices:
 
 def _reference_fdd(vertices_pred, vertices_gt, template, upper_face_map):
     """Reference implementation for FDD metric using numpy."""
-    pred = vertices_pred[:, upper_face_map, :].numpy()  # (T, M, 3)
-    gt = vertices_gt[:, upper_face_map, :].numpy()  # (T, M, 3)
-    template = template[upper_face_map, :].numpy()  # (M, 3)
+    pred = vertices_pred[:, upper_face_map, :].detach().cpu().numpy()  # (T, M, 3)
+    gt = vertices_gt[:, upper_face_map, :].detach().cpu().numpy()  # (T, M, 3)
+    template = template[upper_face_map, :].detach().cpu().numpy()  # (M, 3)
 
     displacements_gt = gt - template  # (T, V, 3)
     displacements_pred = pred - template
@@ -128,9 +128,8 @@ def test_error_on_empty_upper_face_map(self):
 
     def test_error_on_invalid_upper_face_indices(self):
         """Test that an error is raised if upper_face_map has invalid indices."""
-        metric = UpperFaceDynamicsDeviation(template=torch.randn(100, 3), upper_face_map=[98, 99, 100])
         with pytest.raises(ValueError, match="upper_face_map contains invalid vertex indices.*"):
-            metric(torch.randn(10, 50, 3), torch.randn(10, 50, 3))
+            UpperFaceDynamicsDeviation(template=torch.randn(100, 3), upper_face_map=[98, 99, 100])
 
     def test_plot_method(self):
         """Test the plot method of FDD."""